Perceive’s Ergo 2 Supports Transformers

Perceive has launched its second-generation AI SoC, Ergo 2, offering 3–4x higher performance than the previous generation. The new chip targets applications such as security, retail analytics, and visual inspection.

Anand Joshi

Targeting language, speech, and video, Perceive has introduced its second-generation AI SoC: Ergo 2. The new chip provides 3–4x higher performance than the previous generation at an estimated 400 mW (typical). It’s now sampling and will target enterprise-grade camera applications such as security, retail analytics, and visual inspection.

Perceive is a 60-person startup funded by Xperi, which owns media brands such as TiVo and HD Radio. Its first-generation chip, Ergo, arrived in 2020. It consumes an estimated 70 mW to deliver 250 frames per second (fps) on Yolo v3 without needing external DRAM. The chip primarily targets computer vision for cameras, laptops, AR/VR glasses, and similar devices.

The company’s products employ proprietary technology based on information theory to first compress neural networks (NNs) and then accelerate their processing. Perceive claims it can achieve 979 inferences per second (IPS) for ResNet-50 on the ImageNet data set. It has also demonstrated language models as large as Roberta (110 million parameters) running on Ergo 2. The ability to handle Transformer networks at a low power of 400 mW could open a range of new applications such as full speech recognition at the network edge, which is beyond the reach of current chips at similar power.

Although the numbers are impressive, whether the unique framework will keep up with ever-changing neural-network technology is unclear. Perceive might run into trust issues due to the black-boxed nature of compressed NNs. It also must demonstrate that it can handle all customer-required frameworks and operators. The software stack will play a crucial role in the success.